- use a gated activation layer for both attention & convs
- add a relativistic learned position bias. I believe this is similar to the T5 position encodings but it is simpler and learned
- get rid of prepending to the attention matrix - this doesn't really work that well. the model eventually learns to attend one of its heads to these blocks but why not just concat if it is doing that?
This conforms my ConvGnLelu implementation with the generally accepted negative_slope=.2. I have no idea where I got .1. This will break backwards compatibility with some older models but will likely improve their performance when freshly trained. I did some auditing to find what these models might be, and I am not actively using any of them, so probably OK.
